Part 1 of 3 The Partner That Forgets — an arc about what an AI writing partner actually knows about your book.

You open the assistant, and before you can ask it anything useful about chapter nineteen, you write a paragraph explaining who Mara is, why she left the island, and what the reader already knows about the debt. Then you ask your question. Then you get an answer, and it is fine, and it is slightly wrong in a way you can't be bothered to correct, so you take the half of it that works and write the rest yourself.

That paragraph you wrote first has a name in your head, probably something like "setting it up." I want to give it a less forgiving one. It is a briefing, you write a fresh one most sessions, and the reason it exists is the subject of this arc.

The briefing you write before you write

Watch yourself do it once and the shape becomes obvious. A briefing is almost always the same four moves: who is in the scene, what happened recently, what the book is trying to be, and what you want back. You have written some version of it thirty times. You have got good at it — which is its own quiet signal, because nobody gets good at a thing they only do once.

The briefing is not a symptom of a weak assistant. The suggestions that come back are often genuinely good sentences. The problem sits one layer earlier: the assistant answered the briefing beautifully, and the briefing is not the book.

The one-line version

An AI writing partner that has to be briefed is not reasoning about your manuscript. It is reasoning about the summary of your manuscript that you were able to type in ninety seconds while tired.

A summary is not the manuscript

Ask yourself what a good briefing contains, and then ask what it has to leave out. It contains the load-bearing facts as you currently remember them. It leaves out everything you did not think to mention, which includes everything you have forgotten and everything you never consciously decided.

That second category is larger than writers expect. A manuscript is not mostly a set of facts. It is mostly a set of settled questions — how close the narration sits to Mara's head, whether the book does jokes, how long a scene runs before it turns, what kind of sentence you use when something bad is about to happen. You did not decide those at a desk. You decided them across four hundred pages of drafting, one paragraph at a time, and the decisions live in the prose rather than in a notes file.

None of that is briefable. You cannot type "and my narration sits about eight inches behind her left ear except in the flashbacks" because you have never phrased it, even to yourself. You would recognise a violation of it instantly on the page. You just can't hand it over.

What never survives the handover

The middle. Briefings are written from the two ends of a book — the premise you can still recite and the scene you are about to write. Chapters six through fourteen, where most of the actual decisions were made, are represented by a phrase like "she's been on the mainland a while now."

The abandoned versions. You cut a brother in the second draft. He is gone from the manuscript and present in every line that was written while he existed, in the shape of the family, in a silence at a dinner table. A partner working from your briefing will cheerfully suggest a warmth in that house that you spent two drafts removing.

Your own rules. The words you never use. The punctuation you refuse. The fact that you do not write weather at the top of a scene. These are the most identifiably you things in the book and they appear in no briefing anyone has ever written.

Why the last chapter contradicts the eleventh. Continuity is not a list of facts, it is a web of consequences, and the consequences are what you are too close to see. This is precisely the work you would most like handed off.

Why the chat window is a bad place to keep a novel

There is a mechanical reason for all this and it is worth stating plainly, because it is not the assistant being lazy. A general assistant knows what is in front of it. What is in front of it is whatever you pasted, plus whatever it has retained from this conversation. Your novel is not in front of it. A description of your novel is.

So unless you have set something up to carry it, every session starts from roughly zero, and you are the one who has to close the gap — which means the quality of the help you get is capped by how much of your book you can hold in your head and retype at the start of a session. On a good day, with coffee, that is a few thousand words of an eighty-thousand-word manuscript. The partner is working from about four per cent of your book and doing a creditable job on it, and the other ninety-six per cent is a book only you can see.

Try this on your own draft
  • Open the last three sessions you had with any assistant. Count how many begin with a paragraph of set-up.
  • Find a suggestion you rejected. Ask why: was it a bad sentence, or a good sentence for a book slightly different from yours?
  • Pick a decision from the middle of your manuscript that you have never written down anywhere. There will be one within thirty seconds. That decision has never been briefed.

Naming it changes what you are looking for

Most writers file this under "AI isn't there yet for fiction," which is the wrong diagnosis and leads to the wrong shopping. The sentence-level ability is not the constraint. The constraint is that the thing generating sentences has no durable relationship with the book they are supposed to belong to, and no amount of better sentence-generation fixes that.

Which reframes the question you should be asking of any writing partner. Not how good is its prose. Instead: what does it know about my manuscript when I have not just told it, and how did it come to know that?

That is a question with a real answer, and the answers differ enormously. Part two prices what the briefing habit has been quietly doing to the book while you paid it no attention, and part three sets out what a partner has to actually hold before it can be useful at chapter nineteen.